Shrink the install prompt and harden the shipped docs against agent audits#48
Merged
Conversation
…the shipped docs Field-testing showed agents npm-pack the published tarball and read everything in it, so the README section narrating how the prompt "survived AI-agent refusal modes" was itself being quoted as a manipulation signal, and every contradiction between the docs and dist/ was treated as misrepresentation. - Trim the prompt's npmjs.com URL (npm view covers registry verification); aggressive shrinks were refused — the remaining clauses are load-bearing - Move the refusal-mode catalog into field-test/README.md (dev-only, unshipped) and add mode 5: the shipped docs are part of the audit surface - Disclose the opt-in protect command in README + AGENT-INSTALL.md - Scope the "no env vars" payload claim to values; disclose mark-build's env-name stamping - Reconcile the site-UUID guidance (public by design, commit .patchstackrc.json) with the widget's userToken - Soften AGENT-INSTALL's "trust the checklist" and "don't pause" phrasing - Add CLAUDE.md capturing the prompt-is-a-tested-artifact invariants Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Well-structured docs clearly strengthen AI-agent audit defenses. 🎯 Quality: 100% Elite · 📦 Size: Small 📈 This month: Your 213th PR — above team average · Averaging Excellent |
Contributor
Author
|
/review |
mariojgt
approved these changes
Jul 14, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Field-test runs showed hostile agents
npm packthe published tarball and audit everything in it, so this trims the install prompt (the npmjs.com URL was the only clause that proved safely removable — shorter variants were refused) and moves the refusal-mode rationale out of the shipped README intofield-test/README.md, where a new mode 5 documents that the shipped docs are themselves audit surface. It also fixes the doc contradictions agents cited as refusal reasons: the undisclosedprotectcommand is now documented in the README CLI table and AGENT-INSTALL.md, the "no env vars" payload claim is scoped to values withmark-build's env-name stamping disclosed, and the site-UUID guidance no longer says "keep it out of public repos" while the widget ships it in client HTML. A newCLAUDE.mdrecords the prompt-is-a-tested-artifact invariants (prompt.txt ↔ README sync, hostile 3-round gate, no reassurance language). Note: the hostile gate cannot go consistently green until these doc fixes are published, since the harness installs the published package — publish, then re-runnode field-test/run.mjs --persona hostile --rounds 3.🤖 Generated with Claude Code